Re: [Starpu-devel] Suggestions about creating starpu threads

[Fangli Pi <hpcfapix@hlrs.de>]

Hi,

Sorry for the delay of reply.

Fangli

----- Original Message -----
From: "Andra Hugo" <andra.hugo@inria.fr>
To: "Fangli Pi" <hpcfapix@hlrs.de>
Cc: "starpu-devel" <starpu-devel@lists.gforge.inria.fr>, "Samuel Thibault" 
<samuel.thibault@inria.fr>
Sent: Wednesday, May 25, 2016 6:46:11 PM
Subject: Re: Suggestions about creating starpu threads

Hi,

Yes indeed we should first measure if this is actually a problem before going 
into unnecessary modifications.

Sure, I will try to do the assessment of the effect on performance.

Of course as long as we use the tools just for training it should not be a 
problem. However, as far as I know RAPL bases their values on a model (IPC, 
cache misses, frequency etc.) they don't do actual hardware measurement of 
the voltage. So as Sam says, depending on what we executed before hand the 
frequency may change and affect the measurements. But I suppose this can 
happen anyway. Do you do any approximations for this?

Right now I don't have any approximations according to the frequency change, 
but I can have a look at it and try if it is possible.

So these measurements should be desactivated if STARPU_CALIBRATE = 0, right?

According to one of your reviews, I have added one environment to activate 
and deactivate the power measurements, named as STARPU_POWER_PROFILING, which 
determines if ex_starpu_init() is called or not.
 
Andra

> ----- Mail original -----
> > De: "Samuel Thibault" <samuel.thibault@inria.fr>
> > À: "Fangli Pi" <hpcfapix@hlrs.de>
> > Cc: "Andra Hugo" <andra.hugo@inria.fr>, "starpu-devel"
> > <starpu-devel@lists.gforge.inria.fr>
> > Envoyé: Mercredi 25 Mai 2016 16:33:42
> > Objet: Re: Suggestions about creating starpu threads
> > 
> > Fangli Pi, on Wed 25 May 2016 16:30:23 +0200, wrote:
> > > After all, the performance during training is not so important, right?
> > 
> > Hurting the performance of the application is not very important indeed,
> > as long as the performance of the tasks is preserved.
> > 
> > That said, if it really hurts the performance of the application, that
> > could potentially change the energy consumption. If it's not too bad
> > it's fine, but it'd be useful to assess it.
> > 
> > Samuel
> > 
>
-- 
High Performance Computing Center (HLRS)   
University of Stuttgart                    
Nobelstr. 19                               
D-70569 Stuttgart, Germany

phone: ++49-711-685-60442